extern "C" void magma_zlarfxsym(magma_int_t N, magmaDoubleComplex *A, magma_int_t LDA, magmaDoubleComplex *V, magmaDoubleComplex *TAU) { magma_int_t IONE=1; magmaDoubleComplex dtmp; magmaDoubleComplex Z_ZERO = MAGMA_Z_ZERO; //magmaDoubleComplex Z_ONE = MAGMA_Z_ONE; magmaDoubleComplex Z_MONE = MAGMA_Z_NEG_ONE; magmaDoubleComplex Z_HALF = MAGMA_Z_HALF; //magmaDoubleComplex WORK[N]; magmaDoubleComplex *WORK; magma_zmalloc_cpu( &WORK, N ); /* apply left and right on A(st:ed,st:ed)*/ //magma_zlarfxsym(len,A(st,st),LDX,V(st),TAU(st)); /* X = AVtau */ blasf77_zhemv("L",&N, TAU, A, &LDA, V, &IONE, &Z_ZERO, WORK, &IONE); /* je calcul dtmp= X'*V */ dtmp = magma_cblas_zdotc(N, WORK, IONE, V, IONE); /* je calcul 1/2 X'*V*t = 1/2*dtmp*tau */ dtmp = -dtmp * Z_HALF * (*TAU); /* je calcul W=X-1/2VX'Vt = X - dtmp*V */ /* for (j = 0; j < N; j++) WORK[j] = WORK[j] + (dtmp*V[j]); */ blasf77_zaxpy(&N, &dtmp, V, &IONE, WORK, &IONE); /* performs the symmetric rank 2 operation A := alpha*x*y' + alpha*y*x' + A */ blasf77_zher2("L",&N,&Z_MONE,WORK,&IONE,V,&IONE,A,&LDA); magma_free_cpu(WORK); }
inline static void magma_zlarfxsym_v2(magma_int_t n, magmaDoubleComplex *A, magma_int_t lda, magmaDoubleComplex *V, magmaDoubleComplex *TAU, magmaDoubleComplex *work) { /* WORK (workspace) double complex array, dimension N */ magma_int_t ione = 1; magmaDoubleComplex dtmp; magmaDoubleComplex c_zero = MAGMA_Z_ZERO; magmaDoubleComplex c_neg_one= MAGMA_Z_NEG_ONE; magmaDoubleComplex c_half = MAGMA_Z_HALF; /* X = AVtau */ blasf77_zhemv("L",&n, TAU, A, &lda, V, &ione, &c_zero, work, &ione); /* compute dtmp= X'*V */ #if defined(PRECISION_z) || defined(PRECISION_c) dtmp = c_zero; for (magma_int_t j = 0; j < n ; j++) dtmp = dtmp + MAGMA_Z_CNJG(work[j]) * V[j]; //cblas_zdotc_sub(n, work, ione, V, ione, &dtmp); #else dtmp = cblas_zdotc(n, work, ione, V, ione); #endif /* compute 1/2 X'*V*t = 1/2*dtmp*tau */ dtmp = -dtmp * c_half * (*TAU); /* compute W=X-1/2VX'Vt = X - dtmp*V */ blasf77_zaxpy(&n, &dtmp, V, &ione, work, &ione); /* performs the symmetric rank 2 operation A := alpha*x*y' + alpha*y*x' + A */ blasf77_zher2("L", &n, &c_neg_one, work, &ione, V, &ione, A, &lda); }
// -------------------- int main(int argc, char **argv) { TESTING_INIT(); real_Double_t gflops, cpu_time=0, cpu_perf=0, gpu_time, gpu_perf, mgpu_time, mgpu_perf, cuda_time, cuda_perf; double Ynorm, error=0, error2=0, work[1]; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magma_int_t n_local[MagmaMaxGPUs]; magma_int_t N, Noffset, lda, ldda, blocks, lhwork, ldwork, matsize, vecsize; magma_int_t incx = 1; magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 ); magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 ); magmaDoubleComplex *A, *X, *Y, *Ylapack, *Ycublas, *Ymagma, *Ymagma1, *hwork; magmaDoubleComplex_ptr dA, dX, dY; magmaDoubleComplex_ptr d_lA[MagmaMaxGPUs], dwork[MagmaMaxGPUs]; magma_device_t dev; magma_queue_t queues[MagmaMaxGPUs]; magma_int_t status = 0; magma_opts opts; opts.parse_opts( argc, argv ); opts.ngpu = abs( opts.ngpu ); // always uses multi-GPU code double tol = opts.tolerance * lapackf77_dlamch("E"); magma_int_t nb = 64; // required by magmablas_zhemv_mgpu implementation for( dev=0; dev < opts.ngpu; ++dev ) { magma_queue_create( dev, &queues[dev] ); } // currently, tests all offsets in the offsets array; // comment out loop below to test a specific offset. magma_int_t offset = opts.offset; magma_int_t offsets[] = { 0, 1, 31, 32, 33, 63, 64, 65, 100, 200 }; magma_int_t noffsets = sizeof(offsets) / sizeof(*offsets); printf("%% uplo = %s, ngpu %d, block size = %d, offset %d\n", lapack_uplo_const(opts.uplo), (int) opts.ngpu, (int) nb, (int) offset ); printf( "%% BLAS CUBLAS MAGMA 1 GPU MAGMA MGPU Error rel Error rel\n" "%% N offset Gflop/s (msec) Gflop/s (msec) Gflop/s (msec) Gflop/s (msec) to CUBLAS to LAPACK\n" "%%==================================================================================================================\n" ); for( int itest = 0; itest < opts.ntest; ++itest ) { // comment out these two lines & end of loop to test a specific offset for( int ioffset=0; ioffset < noffsets; ioffset += 1 ) { offset = offsets[ioffset]; for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; Noffset = N + offset; lda = Noffset; ldda = magma_roundup( Noffset, opts.align ); // multiple of 32 by default matsize = Noffset*ldda; vecsize = (Noffset-1)*incx + 1; gflops = FLOPS_ZHEMV( N ) / 1e9; blocks = magma_ceildiv( N + (offset % nb), nb ); lhwork = N*opts.ngpu; ldwork = ldda*(blocks + 1); TESTING_MALLOC_CPU( A, magmaDoubleComplex, matsize ); TESTING_MALLOC_CPU( Y, magmaDoubleComplex, vecsize ); TESTING_MALLOC_CPU( Ycublas, magmaDoubleComplex, vecsize ); TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, vecsize ); TESTING_MALLOC_CPU( Ymagma1, magmaDoubleComplex, vecsize ); TESTING_MALLOC_CPU( Ylapack, magmaDoubleComplex, vecsize ); TESTING_MALLOC_PIN( X, magmaDoubleComplex, vecsize ); TESTING_MALLOC_PIN( hwork, magmaDoubleComplex, lhwork ); magma_setdevice( opts.device ); TESTING_MALLOC_DEV( dA, magmaDoubleComplex, matsize ); TESTING_MALLOC_DEV( dX, magmaDoubleComplex, vecsize ); TESTING_MALLOC_DEV( dY, magmaDoubleComplex, vecsize ); // TODO make magma_zmalloc_bcyclic helper function? for( dev=0; dev < opts.ngpu; dev++ ) { n_local[dev] = ((Noffset/nb)/opts.ngpu)*nb; if (dev < (Noffset/nb) % opts.ngpu) n_local[dev] += nb; else if (dev == (Noffset/nb) % opts.ngpu) n_local[dev] += Noffset % nb; magma_setdevice( dev ); TESTING_MALLOC_DEV( d_lA[dev], magmaDoubleComplex, ldda*n_local[dev] ); TESTING_MALLOC_DEV( dwork[dev], magmaDoubleComplex, ldwork ); } ////////////////////////////////////////////////////////////////////////// /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &matsize, A ); magma_zmake_hermitian( Noffset, A, lda ); lapackf77_zlarnv( &ione, ISEED, &vecsize, X ); lapackf77_zlarnv( &ione, ISEED, &vecsize, Y ); /* ===================================================================== Performs operation using CUBLAS =================================================================== */ magma_setdevice( opts.device ); magma_zsetmatrix( Noffset, Noffset, A, lda, dA, ldda, opts.queue ); magma_zsetvector( Noffset, X, incx, dX, incx, opts.queue ); magma_zsetvector( Noffset, Y, incx, dY, incx, opts.queue ); cuda_time = magma_sync_wtime(0); cublasZhemv( opts.handle, cublas_uplo_const(opts.uplo), N, &alpha, dA + offset + offset*ldda, ldda, dX + offset, incx, &beta, dY + offset, incx ); cuda_time = magma_sync_wtime(0) - cuda_time; cuda_perf = gflops / cuda_time; magma_zgetvector( Noffset, dY, incx, Ycublas, incx, opts.queue ); /* ===================================================================== Performs operation using MAGMABLAS (1 GPU) =================================================================== */ magma_setdevice( opts.device ); magma_zsetvector( Noffset, Y, incx, dY, incx, opts.queue ); gpu_time = magma_sync_wtime( opts.queue ); magmablas_zhemv_work( opts.uplo, N, alpha, dA + offset + offset*ldda, ldda, dX + offset, incx, beta, dY + offset, incx, dwork[ opts.device ], ldwork, opts.queue ); gpu_time = magma_sync_wtime( opts.queue ) - gpu_time; gpu_perf = gflops / gpu_time; magma_zgetvector( Noffset, dY, incx, Ymagma1, incx, opts.queue ); /* ===================================================================== Performs operation using MAGMABLAS (multi-GPU) =================================================================== */ magma_zsetmatrix_1D_col_bcyclic( Noffset, Noffset, A, lda, d_lA, ldda, opts.ngpu, nb, queues ); blasf77_zcopy( &Noffset, Y, &incx, Ymagma, &incx ); // workspaces do NOT need to be zero -- set to NAN to prove for( dev=0; dev < opts.ngpu; ++dev ) { magma_setdevice( dev ); magmablas_zlaset( MagmaFull, ldwork, 1, MAGMA_Z_NAN, MAGMA_Z_NAN, dwork[dev], ldwork, opts.queue ); } lapackf77_zlaset( "Full", &lhwork, &ione, &MAGMA_Z_NAN, &MAGMA_Z_NAN, hwork, &lhwork ); mgpu_time = magma_sync_wtime(0); magma_int_t info; info = magmablas_zhemv_mgpu( opts.uplo, N, alpha, d_lA, ldda, offset, X + offset, incx, beta, Ymagma + offset, incx, hwork, lhwork, dwork, ldwork, opts.ngpu, nb, queues ); if (info != 0) { printf("magmablas_zhemv_mgpu returned error %d: %s.\n", (int) info, magma_strerror( info )); } info = magmablas_zhemv_mgpu_sync( opts.uplo, N, alpha, d_lA, ldda, offset, X + offset, incx, beta, Ymagma + offset, incx, hwork, lhwork, dwork, ldwork, opts.ngpu, nb, queues ); if (info != 0) { printf("magmablas_zhemv_sync returned error %d: %s.\n", (int) info, magma_strerror( info )); } mgpu_time = magma_sync_wtime(0) - mgpu_time; mgpu_perf = gflops / mgpu_time; /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { blasf77_zcopy( &Noffset, Y, &incx, Ylapack, &incx ); cpu_time = magma_wtime(); blasf77_zhemv( lapack_uplo_const(opts.uplo), &N, &alpha, A + offset + offset*lda, &lda, X + offset, &incx, &beta, Ylapack + offset, &incx ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; /* ===================================================================== Compute the Difference LAPACK vs. Magma =================================================================== */ Ynorm = lapackf77_zlange( "F", &Noffset, &ione, Ylapack, &Noffset, work ); blasf77_zaxpy( &Noffset, &c_neg_one, Ymagma, &incx, Ylapack, &incx ); error2 = lapackf77_zlange( "F", &Noffset, &ione, Ylapack, &Noffset, work ) / Ynorm; } /* ===================================================================== Compute the Difference Cublas vs. Magma =================================================================== */ Ynorm = lapackf77_zlange( "F", &Noffset, &ione, Ycublas, &Noffset, work ); blasf77_zaxpy( &Noffset, &c_neg_one, Ymagma, &incx, Ycublas, &incx ); error = lapackf77_zlange( "F", &Noffset, &ione, Ycublas, &Noffset, work ) / Ynorm; bool okay = (error < tol && error2 < tol); status += ! okay; if ( opts.lapack ) { printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n", (int) N, (int) offset, cpu_perf, cpu_time*1000., cuda_perf, cuda_time*1000., gpu_perf, gpu_time*1000., mgpu_perf, mgpu_time*1000., error, error2, (okay ? "ok" : "failed") ); } else { printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e --- %s\n", (int) N, (int) offset, cuda_perf, cuda_time*1000., gpu_perf, gpu_time*1000., mgpu_perf, mgpu_time*1000., error, (okay ? "ok" : "failed") ); } /* Free Memory */ TESTING_FREE_CPU( A ); TESTING_FREE_CPU( Y ); TESTING_FREE_CPU( Ycublas ); TESTING_FREE_CPU( Ymagma ); TESTING_FREE_CPU( Ymagma1 ); TESTING_FREE_CPU( Ylapack ); TESTING_FREE_PIN( X ); TESTING_FREE_PIN( hwork ); magma_setdevice( opts.device ); TESTING_FREE_DEV( dA ); TESTING_FREE_DEV( dX ); TESTING_FREE_DEV( dY ); for( dev=0; dev < opts.ngpu; dev++ ) { magma_setdevice( dev ); TESTING_FREE_DEV( d_lA[dev] ); TESTING_FREE_DEV( dwork[dev] ); } fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } // comment out these two lines line & top of loop test a specific offset } // end for ioffset printf( "\n" ); } for( dev=0; dev < opts.ngpu; ++dev ) { magma_queue_destroy( queues[dev] ); } opts.cleanup(); TESTING_FINALIZE(); return status; }
int main(int argc, char **argv) { TESTING_INIT(); real_Double_t gflops, magma_perf, magma_time, cublas_perf, cublas_time, cpu_perf, cpu_time; double magma_error, cublas_error, work[1]; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t N, lda, sizeA, sizeX, sizeY, blocks, ldwork; magma_int_t incx = 1; magma_int_t incy = 1; magma_int_t nb = 64; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 ); magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 ); magmaDoubleComplex *A, *X, *Y, *Ycublas, *Ymagma; magmaDoubleComplex *dA, *dX, *dY, *dC_work; magma_opts opts; parse_opts( argc, argv, &opts ); printf(" N MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA error CUBLAS error\n"); printf("=============================================================================================\n"); for( int i = 0; i < opts.ntest; ++i ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[i]; lda = ((N + 31)/32)*32; sizeA = N*lda; sizeX = N*incx; sizeY = N*incy; gflops = FLOPS_ZHEMV( N ) / 1e9; TESTING_MALLOC( A, magmaDoubleComplex, sizeA ); TESTING_MALLOC( X, magmaDoubleComplex, sizeX ); TESTING_MALLOC( Y, magmaDoubleComplex, sizeY ); TESTING_MALLOC( Ycublas, magmaDoubleComplex, sizeY ); TESTING_MALLOC( Ymagma, magmaDoubleComplex, sizeY ); TESTING_DEVALLOC( dA, magmaDoubleComplex, sizeA ); TESTING_DEVALLOC( dX, magmaDoubleComplex, sizeX ); TESTING_DEVALLOC( dY, magmaDoubleComplex, sizeY ); blocks = (N + nb - 1) / nb; ldwork = lda * (blocks + 1); TESTING_DEVALLOC( dC_work, magmaDoubleComplex, ldwork ); /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &sizeA, A ); magma_zmake_hermitian( N, A, lda ); lapackf77_zlarnv( &ione, ISEED, &sizeX, X ); lapackf77_zlarnv( &ione, ISEED, &sizeY, Y ); /* ===================================================================== Performs operation using CUBLAS =================================================================== */ magma_zsetmatrix( N, N, A, lda, dA, lda ); magma_zsetvector( N, X, incx, dX, incx ); magma_zsetvector( N, Y, incy, dY, incy ); cublas_time = magma_sync_wtime( 0 ); cublasZhemv( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy ); cublas_time = magma_sync_wtime( 0 ) - cublas_time; cublas_perf = gflops / cublas_time; magma_zgetvector( N, dY, incy, Ycublas, incy ); /* ===================================================================== Performs operation using MAGMA BLAS =================================================================== */ magma_zsetvector( N, Y, incy, dY, incy ); magma_time = magma_sync_wtime( 0 ); #if (GPUSHMEM >= 200) magmablas_zhemv2( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy, dC_work, ldwork ); #else magmablas_zhemv( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy ); #endif magma_time = magma_sync_wtime( 0 ) - magma_time; magma_perf = gflops / magma_time; magma_zgetvector( N, dY, incy, Ymagma, incy ); /* ===================================================================== Performs operation using CPU BLAS =================================================================== */ cpu_time = magma_wtime(); blasf77_zhemv( &opts.uplo, &N, &alpha, A, &lda, X, &incx, &beta, Y, &incy ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; /* ===================================================================== Check the result =================================================================== */ blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy ); magma_error = lapackf77_zlange( "M", &N, &ione, Ymagma, &N, work ) / N; blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ycublas, &incy ); cublas_error = lapackf77_zlange( "M", &N, &ione, Ycublas, &N, work ) / N; printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e\n", (int) N, magma_perf, 1000.*magma_time, cublas_perf, 1000.*cublas_time, cpu_perf, 1000.*cpu_time, magma_error, cublas_error ); TESTING_FREE( A ); TESTING_FREE( X ); TESTING_FREE( Y ); TESTING_FREE( Ycublas ); TESTING_FREE( Ymagma ); TESTING_DEVFREE( dA ); TESTING_DEVFREE( dX ); TESTING_DEVFREE( dY ); TESTING_DEVFREE( dC_work ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return 0; }
int main(int argc, char **argv) { TESTING_INIT(); const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; const magma_int_t ione = 1; real_Double_t atomics_perf=0, atomics_time=0; real_Double_t gflops, magma_perf=0, magma_time=0, cublas_perf, cublas_time, cpu_perf, cpu_time; double magma_error=0, atomics_error=0, cublas_error, work[1]; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t N, lda, ldda, sizeA, sizeX, sizeY, blocks, ldwork; magma_int_t incx = 1; magma_int_t incy = 1; magma_int_t nb = 64; magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 ); magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 ); magmaDoubleComplex *A, *X, *Y, *Yatomics, *Ycublas, *Ymagma; magmaDoubleComplex_ptr dA, dX, dY, dwork; magma_int_t status = 0; magma_opts opts; opts.parse_opts( argc, argv ); double tol = opts.tolerance * lapackf77_dlamch("E"); printf("%% uplo = %s\n", lapack_uplo_const(opts.uplo) ); printf("%% N MAGMA Gflop/s (ms) Atomics Gflop/s CUBLAS Gflop/s CPU Gflop/s MAGMA error Atomics CUBLAS\n"); printf("%%=====================================================================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; lda = N; ldda = magma_roundup( N, opts.align ); // multiple of 32 by default sizeA = N*lda; sizeX = N*incx; sizeY = N*incy; gflops = FLOPS_ZHEMV( N ) / 1e9; TESTING_MALLOC_CPU( A, magmaDoubleComplex, sizeA ); TESTING_MALLOC_CPU( X, magmaDoubleComplex, sizeX ); TESTING_MALLOC_CPU( Y, magmaDoubleComplex, sizeY ); TESTING_MALLOC_CPU( Yatomics, magmaDoubleComplex, sizeY ); TESTING_MALLOC_CPU( Ycublas, magmaDoubleComplex, sizeY ); TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, sizeY ); TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*N ); TESTING_MALLOC_DEV( dX, magmaDoubleComplex, sizeX ); TESTING_MALLOC_DEV( dY, magmaDoubleComplex, sizeY ); blocks = magma_ceildiv( N, nb ); ldwork = ldda*blocks; TESTING_MALLOC_DEV( dwork, magmaDoubleComplex, ldwork ); magmablas_zlaset( MagmaFull, ldwork, 1, MAGMA_Z_NAN, MAGMA_Z_NAN, dwork, ldwork ); magmablas_zlaset( MagmaFull, ldda, N, MAGMA_Z_NAN, MAGMA_Z_NAN, dA, ldda ); /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &sizeA, A ); magma_zmake_hermitian( N, A, lda ); // should not use data from the opposite triangle -- fill with NAN to check magma_int_t N1 = N-1; if ( opts.uplo == MagmaUpper ) { lapackf77_zlaset( "Lower", &N1, &N1, &MAGMA_Z_NAN, &MAGMA_Z_NAN, &A[1], &lda ); } else { lapackf77_zlaset( "Upper", &N1, &N1, &MAGMA_Z_NAN, &MAGMA_Z_NAN, &A[lda], &lda ); } lapackf77_zlarnv( &ione, ISEED, &sizeX, X ); lapackf77_zlarnv( &ione, ISEED, &sizeY, Y ); /* ===================================================================== Performs operation using CUBLAS =================================================================== */ magma_zsetmatrix( N, N, A, lda, dA, ldda ); magma_zsetvector( N, X, incx, dX, incx ); magma_zsetvector( N, Y, incy, dY, incy ); magmablasSetKernelStream( opts.queue ); // opts.handle also uses opts.queue cublas_time = magma_sync_wtime( opts.queue ); #ifdef HAVE_CUBLAS cublasZhemv( opts.handle, cublas_uplo_const(opts.uplo), N, &alpha, dA, ldda, dX, incx, &beta, dY, incy ); #else magma_zhemv( opts.uplo, N, alpha, dA, 0, ldda, dX, 0, incx, beta, dY, 0, incy, opts.queue ); #endif cublas_time = magma_sync_wtime( opts.queue ) - cublas_time; cublas_perf = gflops / cublas_time; magma_zgetvector( N, dY, incy, Ycublas, incy ); /* ===================================================================== Performs operation using CUBLAS - using atomics =================================================================== */ #ifdef HAVE_CUBLAS cublasSetAtomicsMode( opts.handle, CUBLAS_ATOMICS_ALLOWED ); magma_zsetvector( N, Y, incy, dY, incy ); // sync on queue doesn't work -- need device sync or use NULL stream -- bug in CUBLAS? atomics_time = magma_sync_wtime( NULL /*opts.queue*/ ); cublasZhemv( opts.handle, cublas_uplo_const(opts.uplo), N, &alpha, dA, ldda, dX, incx, &beta, dY, incy ); atomics_time = magma_sync_wtime( NULL /*opts.queue*/ ) - atomics_time; atomics_perf = gflops / atomics_time; magma_zgetvector( N, dY, incy, Yatomics, incy ); cublasSetAtomicsMode( opts.handle, CUBLAS_ATOMICS_NOT_ALLOWED ); #endif /* ===================================================================== Performs operation using MAGMABLAS =================================================================== */ #ifdef HAVE_CUBLAS magma_zsetvector( N, Y, incy, dY, incy ); magma_time = magma_sync_wtime( opts.queue ); if ( opts.version == 1 ) { magmablas_zhemv_work( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy, dwork, ldwork, opts.queue ); } else { // non-work interface (has added overhead) magmablas_zhemv( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy ); } magma_time = magma_sync_wtime( opts.queue ) - magma_time; magma_perf = gflops / magma_time; magma_zgetvector( N, dY, incy, Ymagma, incy ); #endif /* ===================================================================== Performs operation using CPU BLAS =================================================================== */ cpu_time = magma_wtime(); blasf77_zhemv( lapack_uplo_const(opts.uplo), &N, &alpha, A, &lda, X, &incx, &beta, Y, &incy ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; /* ===================================================================== Check the result =================================================================== */ blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ycublas, &incy ); cublas_error = lapackf77_zlange( "M", &N, &ione, Ycublas, &N, work ) / N; #ifdef HAVE_CUBLAS blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Yatomics, &incy ); atomics_error = lapackf77_zlange( "M", &N, &ione, Yatomics, &N, work ) / N; blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy ); magma_error = lapackf77_zlange( "M", &N, &ione, Ymagma, &N, work ) / N; #endif bool okay = (magma_error < tol && cublas_error < tol && atomics_error < tol); status += ! okay; printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %8.2e %s\n", (int) N, magma_perf, 1000.*magma_time, atomics_perf, 1000.*atomics_time, cublas_perf, 1000.*cublas_time, cpu_perf, 1000.*cpu_time, magma_error, cublas_error, atomics_error, (okay ? "ok" : "failed")); TESTING_FREE_CPU( A ); TESTING_FREE_CPU( X ); TESTING_FREE_CPU( Y ); TESTING_FREE_CPU( Ycublas ); TESTING_FREE_CPU( Yatomics ); TESTING_FREE_CPU( Ymagma ); TESTING_FREE_DEV( dA ); TESTING_FREE_DEV( dX ); TESTING_FREE_DEV( dY ); TESTING_FREE_DEV( dwork ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } opts.cleanup(); TESTING_FINALIZE(); return status; }
int main(int argc, char **argv) { TESTING_INIT(); real_Double_t gflops, magma_perf, magma_time, cublas_perf, cublas_time, cpu_perf, cpu_time; double magma_error, cublas_error, work[1]; magma_int_t ione = 1; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t N, lda, ldda, sizeA, sizeX, sizeY, blocks, ldwork; magma_int_t incx = 1; magma_int_t incy = 1; magma_int_t nb = 64; magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE; magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 ); magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 ); magmaDoubleComplex *A, *X, *Y, *Ycublas, *Ymagma; magmaDoubleComplex *dA, *dX, *dY, *dwork; magma_int_t status = 0; magma_opts opts; parse_opts( argc, argv, &opts ); double tol = opts.tolerance * lapackf77_dlamch("E"); printf("uplo = %s\n", lapack_uplo_const(opts.uplo) ); if ( opts.uplo == MagmaUpper ) { printf("** for uplo=MagmaUpper, magmablas_zhemv simply calls cublas_zhemv.\n"); } printf(" N MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA error CUBLAS error\n"); printf("=============================================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { N = opts.nsize[itest]; lda = N; ldda = ((N + 31)/32)*32; sizeA = N*lda; sizeX = N*incx; sizeY = N*incy; gflops = FLOPS_ZHEMV( N ) / 1e9; TESTING_MALLOC_CPU( A, magmaDoubleComplex, sizeA ); TESTING_MALLOC_CPU( X, magmaDoubleComplex, sizeX ); TESTING_MALLOC_CPU( Y, magmaDoubleComplex, sizeY ); TESTING_MALLOC_CPU( Ycublas, magmaDoubleComplex, sizeY ); TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, sizeY ); TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*N ); TESTING_MALLOC_DEV( dX, magmaDoubleComplex, sizeX ); TESTING_MALLOC_DEV( dY, magmaDoubleComplex, sizeY ); blocks = (N + nb - 1) / nb; ldwork = ldda*blocks; TESTING_MALLOC_DEV( dwork, magmaDoubleComplex, ldwork ); magmablas_zlaset( MagmaFull, ldwork, 1, MAGMA_Z_NAN, MAGMA_Z_NAN, dwork, ldwork ); magmablas_zlaset( MagmaFull, ldda, N, MAGMA_Z_NAN, MAGMA_Z_NAN, dA, ldda ); /* Initialize the matrix */ lapackf77_zlarnv( &ione, ISEED, &sizeA, A ); magma_zmake_hermitian( N, A, lda ); lapackf77_zlarnv( &ione, ISEED, &sizeX, X ); lapackf77_zlarnv( &ione, ISEED, &sizeY, Y ); /* ===================================================================== Performs operation using CUBLAS =================================================================== */ magma_zsetmatrix( N, N, A, lda, dA, ldda ); magma_zsetvector( N, X, incx, dX, incx ); magma_zsetvector( N, Y, incy, dY, incy ); cublas_time = magma_sync_wtime( 0 ); cublasZhemv( handle, cublas_uplo_const(opts.uplo), N, &alpha, dA, ldda, dX, incx, &beta, dY, incy ); cublas_time = magma_sync_wtime( 0 ) - cublas_time; cublas_perf = gflops / cublas_time; magma_zgetvector( N, dY, incy, Ycublas, incy ); /* ===================================================================== Performs operation using MAGMABLAS =================================================================== */ magma_zsetvector( N, Y, incy, dY, incy ); //magma_zprint_gpu( ldda, blocks, dwork, ldda ); magma_time = magma_sync_wtime( 0 ); magmablas_zhemv_work( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy, dwork, ldwork ); // TODO provide option to test non-work interface //magmablas_zhemv( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy ); magma_time = magma_sync_wtime( 0 ) - magma_time; magma_perf = gflops / magma_time; magma_zgetvector( N, dY, incy, Ymagma, incy ); //magma_zprint_gpu( ldda, blocks, dwork, ldda ); /* ===================================================================== Performs operation using CPU BLAS =================================================================== */ cpu_time = magma_wtime(); blasf77_zhemv( lapack_uplo_const(opts.uplo), &N, &alpha, A, &lda, X, &incx, &beta, Y, &incy ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; /* ===================================================================== Check the result =================================================================== */ blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy ); magma_error = lapackf77_zlange( "M", &N, &ione, Ymagma, &N, work ) / N; blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ycublas, &incy ); cublas_error = lapackf77_zlange( "M", &N, &ione, Ycublas, &N, work ) / N; printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n", (int) N, magma_perf, 1000.*magma_time, cublas_perf, 1000.*cublas_time, cpu_perf, 1000.*cpu_time, magma_error, cublas_error, (magma_error < tol && cublas_error < tol ? "ok" : "failed")); status += ! (magma_error < tol && cublas_error < tol); TESTING_FREE_CPU( A ); TESTING_FREE_CPU( X ); TESTING_FREE_CPU( Y ); TESTING_FREE_CPU( Ycublas ); TESTING_FREE_CPU( Ymagma ); TESTING_FREE_DEV( dA ); TESTING_FREE_DEV( dX ); TESTING_FREE_DEV( dY ); TESTING_FREE_DEV( dwork ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }